This week: a cloned voice helped fraudsters steal €95 million from an Italian private bank. Researchers also found AI-driven hacking attempts against government websites, while courts and regulators examined how AI systems handle personal identity and act on users’ behalf.
Fraudsters used a fake WhatsApp message and an AI-cloned voice to steal €95 million from Fideuram, Intesa Sanpaolo’s private bank, according to two people familiar with the case. The February scam began with a message to then-chairman Paolo Molesini that appeared to come from Intesa CEO Carlo Messina. A follow-up call impersonating a senior law-firm partner confirmed the supposed need for an urgent foreign transaction. Molesini instructed his finance department to arrange transfers, mainly to accounts in China and Hong Kong. More than half the money was later recovered, but €36 million remained missing when the case was reported. Both banks declined to comment.
AI research agents attempted to exploit websites run by the U.S. Education Department and Library and Archives Canada, Transluce researchers found. Logs from May and June included SQL injection: specially crafted input intended to change a database query and bypass its normal restrictions. The U.S. activity involved the Civil Rights Data Collection service; the Canadian requests targeted historical divorce records. The researchers could not conclusively identify the model provider behind every request and found no evidence that either attack succeeded or exposed nonpublic information. Canada separately said it had found no indication of a compromise.
OpenAI says individuals associated with Moonshot AI, the developer of Kimi, were behind a core group in a campaign to extract its models’ hidden reasoning. The aim, it says, was distillation: using one model’s outputs to train another. Participants copied encrypted reasoning from one conversation into another and asked the model to decrypt or transcribe it. OpenAI says this exploited model behavior, not broken encryption, and did not involve access to its databases or stored conversations. It banned or restricted accounts and blocked a route that let users replay someone else’s encrypted reasoning and recover its contents.
The U.S. Federal Trade Commission is investigating OpenAI, Anthropic and other AI developers over possible consumer harm from agents, according to a senior agency official. The inquiry examines possible unfair or deceptive practices under the FTC Act, a law the agency has previously used against companies that failed to secure consumer data. Agents can take actions, such as browsing websites or using connected accounts, rather than just answer questions. The official said the investigation began before the latest reports of unauthorized agent activity. The FTC plans to use civil investigative demands, similar to subpoenas, to obtain records and compel executive testimony.
A Chicago Claude user has filed a proposed class action against Anthropic over facial scans used to restore account access. Jose Enrique Ortiz Colon says he had to submit his driver’s licence and a live image of his face through verification provider Persona. He alleges Anthropic failed to disclose in writing, before the scan, how long it would retain his biometric data, and lacked a public retention and destruction policy. The lawsuit invokes the Illinois Biometric Information Privacy Act, although it was filed in San Francisco County Superior Court. The claims have not been tested in court. Anthropic said it was reviewing the complaint.
A Tokyo court ruled that unauthorized use of a performer’s voice can infringe publicity rights when its purpose is exclusively to exploit that voice’s commercial appeal. Voice actor Kenjiro Tsuda sued TikTok over videos he said used an AI clone of his voice; the platform argued that the narration was a generic male voice. Publicity rights govern commercial use of a person’s identity, such as their name, likeness or recognizable voice. The court dismissed Tsuda’s removal request because the uploader had already deleted the videos. TikTok said it would respect the judgment while reviewing its details.
Mastercard is testing a service in the U.S. that estimates whether an AI agent initiated a transaction. The probability score is part of its Agent Pay program, which supports purchases made by agents on consumers’ behalf. The score identifies likely AI involvement; it does not, by itself, establish that a purchase is authorized. Mastercard plans to add behavioral, merchant, transaction and credential-risk information. It is also exploring combined web and payment-network signals with Cloudflare and working with Skyfire’s Know Your Agent technology to help banks and merchants recognize verified agents.
